Conclusion of Deep Learning and Questions

(28th-October-2020)

• DNN has achieved great results - a large margin in performance ratio of existing methods - still room for performance improvement

• There was not a particularly big theoretical progress (now) - is pre-training necessary? - CNN has remained almost unchanged since the latter half of the 1980s - that it turned out "It can be done if you do it"

• Improvement in computer performance · It is unchanged from the old days that know-how to derive performance is necessary

• Learning with large-scale data by large-scale NN - I want to improve recognition performance → NN multilayered → excessive learning risk → increase learning data amount → calculation performance required

• Expectations for Representation / feature learning - semi-supervised / transfer / self-taught learning

Questions.

1. Design Restricted Boltzmann Machine (RBM) for specific domain model.

2. Explain Convolutional Neural Network (CNN).

3. Why Deep NN is getting popular than Shallow NN?